Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Forum Math. 24 (2012), 445 – 470 DOI 10.1515/FORM.2011.075 Forum Mathematicum © de Gruyter 2012 V -variable fractals: dimension results Michael Barnsley, John E. Hutchinson and Örjan Stenflo Communicated by Frank Duzaar Abstract. The families of V -variable fractals for V D 1; 2; 3; : : : , together with their natural probability distributions, interpolate between the corresponding families of random homogeneous fractals and of random recursive fractals. We investigate certain random V V matrices associated with these fractals and use them to compute the almost sure Hausdorff dimension of V -variable fractals satisfying the uniform open set condition. Keywords. V -variable fractals, Hausdorff dimension. 2010 Mathematics Subject Classification. Primary 28A80; secondary 37H99, 60G57, 60J05. 1 1.1 Introduction Overview In this paper we begin the study of analysis on V -variable fractals by computing their Hausdorff dimension, under the assumption of an open set condition. In order to make the paper self-contained we include an informal discussion of some simple examples of V -variable fractals and their relationship to other notions of random fractals. For more details see [4]. The key idea is to code up relevant information as a product of random V V matrices, one for each level in the construction of a generic V -variable fractal. The almost sure growth rate of the norms of these products can be obtained from a variant of the Furstenberg–Kesten theorem for products of random matrices. Necks are defined in Definition 5.3 and used in (5.11) and (5.16) to construct an appropriate comparison measure on a generic V -variable fractal, and then to bound the local mass growth rate of in (5.17) and (5.21). This work was partially supported by the Australian Research Council and carried out at the Australian National University. Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 446 M. Barnsley, J. E. Hutchinson and Ö. Stenflo 1.2 Background Fractal sets generated by a single (contractive) iterated function system [IFS], and random fractal sets generated from a family of such IFSs together with a probability distribution on this family, are studied as mathematical models of disordered systems. In this latter setting the most commonly studied random fractals are random recursive fractals and random homogeneous fractals. In particular, Hausdorff, walk and spectral dimensions have been computed in special cases. See [7, 8, 14] for general background on fractals and random fractals, including their applications, and see [18, 15, 6, 10, 11, 17, 12, 13, 1, 16, 21] and the references therein for the study of the various dimensions and other analytic properties of fractals. The classes of V -variable fractals, together with their natural probability distributions, are defined in the next section. For V D 1; 2; : : : , they interpolate between recursive and homogeneous fractals, and similarly for the random versions in each case. They have some initially surprising properties as noted in Section 1.4. In particular, they correspond to the elements of the attractor of a single deterministic IFS, operating not on Rn but on the metric space of V -tuples of compact subsets of Rn . Their natural probability distribution can be obtained from the unit mass measure on this deterministic IFS, and so they can be rapidly generated by a “chaos game” or Monte Carlo Markov Chain [MCMC] algorithm. 1.3 Preliminary notation Fix .F ; P / where F is a collection of (contractive) IFSs operating on Rn and P is a probability distribution on F . From this data, as sketched in the model example in Section 2, one constructs a pair .K1 ; K1 / where K1 is the class of recursive fractals corresponding to .F ; P / and K1 is a natural probability distribution on K1 , see [6, 10, 17]. One also constructs a corresponding pair .K1 ; K1 / where K1 is the corresponding class of homogeneous fractals and K1 is the natural probability distribution on K1 , see [12]. For each natural number V we also construct the family KV of V -variable fractals together with a natural probability distribution KV on KV . These families .KV ; KV / interpolate between the previous two classes .K1 ; K1 / (the V D 1 case) and .K1 ; K1 / (the limit case as V ! 1). Each class of V -variable random fractals, with its probability distribution, has the surprising property that it can be obtained from the attractor of a single IFS operating on V -tuples of compact subsets of Rn . Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM V -variable fractals: dimension results 447 For the general development of V -variable fractals see [4], and for other examples and discussion see [3, 2]. 1.4 Properties of V -variable fractals We are interested in V -variable fractals for the following reasons, see [4, 3]. (i) Families of V -variable fractals, together with their probability distributions, interpolate between random homogeneous fractals and random recursive fractals. (ii) Certain families of functions .ˆa ; a 2 AV ; PV / and .F a ; a 2 AV ; PV / together with the probability distribution PV on AV are IFSs with probabilities in the standard sense, the first acting on V and the other on C.R2 /V . (See (2.1) and (2.2), or Section (4.1), for the relevant definitions.) Their measure attractors on V and C .R2 /V respectively, projected in any one of the V component directions, give the collection of random V -variable fractal trees and sets together with their natural probability distributions. (iii) The chaos game for these IFSs can be used to generate a sample of V -variable fractals, whose empirical distribution approaches the probability distribution on V -variable fractals as the sample size approaches infinity. (iv) Analogous results apply to V -variable fractal measures under weak local contractive conditions as opposed to strict global contractivity. Such conditions are natural, for example, in modelling stochastic processes where individual sample paths may be bounded, but there is no uniform bound. In this case one has an IFS operating on a non-locally compact state space. But the chaos game result can be extended to this setting. (v) By taking large V the chaos game gives a fast forward process for the generation of a sample of fractals approximating random recursive fractals and their probability distribution. This is of practical interest, since normally one builds individual examples of random recursive fractals by a computationally expensive backward process. 2 2.1 Sierpinski gaskets, a model example Recursive and homogeneous gaskets Let F D .R2 I f1 ; f2 ; f3 / be the IFS consisting of three contraction maps on R2 , each with contraction ratio 1=2, and having fixed points that are the vertices of an equilateral triangle T of unit diameter. Let G D .R2 I f1 ; f2 ; f3 / be the IFS Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 448 M. Barnsley, J. E. Hutchinson and Ö. Stenflo consisting of three contraction maps, each with contraction ratio 1=3, and having the same fixed points as the corresponding maps in F . The attractors of F and G are denoted by SF and SG respectively, see Figure 1. Figure 1. Prefractal approximations to the attractors SF and SG respectively. Consider F D ¹F; Gº, together with the probability distribution P D ¹1=2; 1=2º on F . Let denote the set of labelled trees which are rooted, 3-branching, and infinite, where the label at each node is either F or G; see Section 3.2. A modified Sierpinski gaket K ! can be generated from each ! 2 , see Figure 2. Such fractals K ! are examples of recursive fractals. If the nodes of ! at each fixed level are the same but may vary from level to level, then K ! is called a homogeneous fractal. It will also be convenient to consider finite, level k, labelled trees for each k 0; see Figure 2. Such trees correspond to level k prefractals. Figure 2. Level 3 recursive prefractals that are 1, 2 and 3-variable respectively, together with the corresponding finite labelled trees. Each vertex of the triangle T is labelled according to the function which fixes it. The labelling convention is clear: the 3 tree labels above each node when read from left to right correspond to applying the corresponding functions onto the bottom left, top and bottom right respectively of the relevant triangle. Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM V -variable fractals: dimension results 449 Fractals such as SF and SG have the following properties: (i) Spatial self similarity: loosely expressed, at each fixed “scale” the component parts are equivalent up to simple transformations, for example, translations in the case here. (ii) Scale self similarity: the equivalence class at each scale is the same. Homogeneous fractals have spatial self similarity but, generically, do not have scale self similarity. Recursive fractals generically have neither spatial nor scale self similarity. Both classes have statistical self similarity if one imposes suitable probability distributions on their construction. If the labels of the tree ! are chosen in an iid manner according to some probability distribution P , except that in the homogeneous case all labels at each fixed level are same, then one obtains random recursive fractals and random homogeneous fractals respectively. 2.2 V -variable gaskets Now assume that, at each level of ! 2 , the subtrees rooted at that level have the property that they belong to at most V distinct isomorphism classes. In this case, ! is said to be a V -variable labelled tree, and the fractal K ! is said to be a V -variable gasket or fractal. Such fractals have a form of partial spatial self similarity. Similarly, define V -variable finite labelled trees and V -variable prefractals as in Figure 2. Notice that homogeneous fractals and 1-variable fractals are the same. Let V denote the class of V -variable trees ! corresponding to F , and let KV denote the corresponding class of V -variable fractals K ! . As is illustrated in Figure 3, there are natural maps ˆa from V -tuples of infinite, respectively level k, labelled trees .!1 ; : : : ; !V / to V -tuples of infinite, respectively level k C 1, labelled trees .!10 ; : : : ; !V0 /, respectively. That is, .!10 ; : : : ; !V0 / D ˆa .!1 ; : : : ; !V /: (2.1) The label at the base or level 0 node of each !v0 is determined by ˆa and is either F or G. The three subtrees of each !v0 rooted at level one are also determined by ˆa and taken from the set ¹!1 ; : : : ; !V º, possibly with repetition. The maps ˆa are described by V .M C 1/ arrays a, where M D 3 and V D 4 in Figure 3. In general, M is the maximum cardinality of the set of functions in each IFS from the family F . Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 450 M. Barnsley, J. E. Hutchinson and Ö. Stenflo Figure 3. A V D 4 example. The map ˆa0 is described by the array a0 . It is also described by the labels in the level 0 (bottom) boxes, and the network of lines between these boxes and the level 1 (middle) boxes. For example, row 3 of a0 is G; 2; 4; 2. It contains the information that the third component of ˆa0 .!1 ; !2 ; !3 ; !4 / is the tree whose root node is labelled G, and whose three subtrees, rooted at the next level, are !2 , !4 and !2 respectively. This information is also provided by the facts that box 3 at level 0 contains the symbol G, and the three lines from this box in the order left to right connect to boxes 2, 4 and 2 at level 1. Similar remarks apply to ˆa1 . Corresponding to the family F let AV D the set of all such arrays (indices) a; A1 V D ¹a0 a1 : : : ak : : : W ai 2 AV for all i º: (2.2) The maps F a act on V -tuples of compact subsets of Rn in an analogous manner, see Figure 4. If a D a0 a1 : : : ak : : : 2 A1 V , then .!1a ; : : : ; !Va / WD lim ˆa0 ı ı ˆak 1 .!10 ; : : : ; !V0 /; k!1 .K1a ; : : : ; KVa / WD lim F a0 ı ı F ak 1 .K10 ; : : : ; KV0 /: (2.3) k!1 The limits exist and are independent of .!10 ; : : : ; !V0 / and .K10 ; : : : ; KV0 / respectively. Up to level k, the labelled trees .!1a ; : : : ; !Va / depend only on the maps ˆa0 ; : : : ; ˆak 1 , see Figure 4 where k D 2. Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM V -variable fractals: dimension results 451 Figure 4. The map F a1 followed by F a0 act on C .R2 /4 , the space of 4-tuples of compact subsets of R2 , in a manner which is encoded in the corresponding arrays a0 and a1 . For example, row 3 of a0 is G; 2; 4; 2 and contains the information that, for any .K1 ; K2 ; K3 ; K4 /, component 3 of F .K1 ; K2 ; K3 ; K4 / will be g1 .K2 / [ g2 .K4 / [ g3 .K2 /. Note a0 and a1 are as in Figure 3. 2.3 Symbolic representation There are two ways of representing V -variable fractals. First, any V -variable fractal K can be represented by a labelled tree ! 2 as in Figure 2. This does not utilise the V -variability. Second, and more useful here, the fractal sets Kva in (2.3) are also described by the infinite sequence a D a0 a1 : : : ak : : :, often called an address for the V -tuple K a . The map a 7! K a is typically many-to-one. The labelled trees !va and the fractal sets Kva are V -variable, and every V -variable tree and fractal set can be obtained in this manner. Not only is .!1a ; : : : ; !Va / a V -tuple of V -variable trees, but it satisfies the stronger condition that, for each k, there are at most V isomorphism classes of trees rooted at level k, chosen from all the !1a ; : : : ; !Va taken together. A similar remark applies to .K1a ; : : : ; KVa /. 2.4 Random V -variable gaskets So far we have not required a probability distribution on the class V of V -variable trees or on the class KV of V -variable fractals. However, there is a natural probability distribution PV on AV that is inherited from the probability distribution P on the family of IFSs F D ¹F; Gº. This induces a probability distribution V V V on A1 V , thence from (2.3) on .V / and .KV / , and thence by projection on V and KV . More precisely, all components of a 2 AV are chosen independently; those in the first column from F according to P and the remaining entries from ¹1; : : : ; V º according to the uniform distribution. The resulting probability distribution PV on AV induces a probability distribution on A1 V , obtained by choosing the ele- Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 452 M. Barnsley, J. E. Hutchinson and Ö. Stenflo ments ak of the sequence a D a0 a1 : : : ak : : : in an iid manner according to PV . V The mapping A1 V ! given by (2.3) now induces a probability distribution on V V .V / . The projected distribution on V in any coordinate direction is independent of the coordinate direction, and similarly for KV . The probability distribution on 1 obtained in this manner is the same as the random homogeneous distribution. The distribution on V . / converges to the random recursive distribution on as V ! 1. See [4] for details. 2.5 Flow matrices Recall that the three fixed points of f1 , f2 and f3 are also the fixed points of g1 , g2 and g3 ; namely, the vertices of an equilateral triangle T of unit diameter. Suppose K D K ! is a recursive Sierpinski triangle with labelled tree ! 2 . The 3k scaled triangles in the k-level prefractal approximation provide an efficient cover of K for large k, see Figure 2 where k D 3, and Figure 4 where k D 2. In order to study the Hausdorff measure H ˛ .K/, it is natural to consider X S.!; k; ˛/ WD jTm j˛ : ¹m2!WjmjDkº Here, on the right side, we are summing the diameters to the power ˛ of the 3k triangles in the level k prefractal for K, see also (4.4). As in Figure 5, let .K10 ; : : : ; KV0 / D F a .K1 ; : : : ; KV /, where .K1 ; : : : ; KV / has labelled trees ! D .!1 ; : : : ; !V / and .K10 ; : : : ; KV0 / has labelled trees !0 D .!10 ; : : : ; !V0 /. Setting S .!; k; ˛/ D .S.!1 ; k; ˛/; : : : ; S.!V ; k; ˛//, one can see by examining Figure 5 that S .!0 ; k; ˛/ D M a .˛/S .!; k 1; ˛/; (2.4) where, in general, M a .˛/ is the V V flow matrix constructed from a as in Definition 4.2. See also Proposition 4.1. If !a D .!1a ; : : : ; !Va / and .K1a ; : : : ; KVa / are as in (2.3), it follows on iterating (2.4) that S .!a ; k; ˛/ D M a0 ı ı M ak 1 1: The vector 1 is a vector of units, see also Proposition 4.3. For a matrix M , let kM k be obtained by summing the absolute values of all components. It is now plausible that there is a unique d such that kM a0 ı ı M ak 1 k should diverge exponentially fast to C1 a.s. for ˛ < d , decay exponentially fast to 0 a.s. for ˛ > d , and that this d should be the a.s. Hausdorff dimension of K1a ; : : : ; KVa . We show this is the case in the main theorem, see Sections 4 and 5. In Section 6 we compute some examples. Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM V -variable fractals: dimension results 453 Figure 5. The first row represents fractals .K1 ; : : : ; K4 /, or prefractals of level k 1, corresponding to labelled trees ! D .!1 ; : : : ; !4 /. The second row represents .K10 ; : : : ; K40 / D F a .K1 ; : : : ; K4 / with labelled trees !0 D .!10 ; : : : ; !40 /. If the vth component of S .!; k 1; ˛/ isthe k 1 level approximation to H ˛ .Kv /, ˛ ˛ then, in the case here, S.!30 ; k; ˛/ D 13 S.!2 ; k 1; ˛/ C 2 13 S.!3 ; k 1; ˛/. Similarly, S .!0 ; k; ˛/ D M a .˛/S .!; k 1; ˛/. Results for walk and spectral dimensions of V -variable fractals, and their properties, have been obtained in work by Uta Freiberg, Ben Hambly and Hutchinson. 3 Notation and assumptions The diameter of a set A is denoted by jAj. The Hausdorff measure of A in the dimension ˛ is denoted by H ˛ .A/. The Hausdorff dimension of A is denoted by dimH .A/. 3.1 A family of contractive IFSs Fix a family F D ¹F W 2 ƒº of IFSs acting on the space Rn , where we have < 1, F D .Rn I f1 ; : : : ; fM /, each fm W Rn ! Rn with Lipschitz constant rm and ƒ may be infinite. Fix a probability distribution P on ƒ, or equivalently on F . Assume 0 < rmin rm rmax < 1; M WD max M < 1: 2ƒ (3.1) We usually further assume the fm W Rn ! Rn are similitudes. That is, jfm .x/ fm .y/j D rm jx yj 8x; y 2 Rn : (3.2) Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 454 M. Barnsley, J. E. Hutchinson and Ö. Stenflo In this case we will also assume that F satisfies a uniform open set condition. That is, there exists a non-empty open set O such that M [ fm .O/ O; fm .O/ \ fn .O/ D ; if m ¤ n and 2 ƒ: (3.3) mD1 We note explicitly when (3.2) and (3.3) apply. 3.2 Labelled trees We characterise the set of labelled trees !, corresponding to the family F of IFSs, as follows: An (unlabelled) tree is a set ! of finite sequences m D m1 : : : mk where each mj 2 ¹1; : : : ; M º, together with the empty sequence ;, and which is closed under the operation of taking initial segments. Sequences m 2 ! are called nodes. (Note that the branching number at each node is M .) The level jmj of the node m D m1 : : : mk is defined to be jmj D k, while j;j D 0. Some, but not all, such trees, correspond to one or more labelled trees. More precisely: A labelled tree, corresponding to the family of IFSs F , is a tree ! together with a map from the set of nodes of ! into ƒ. This map is also denoted by !. If m 2 ! and !.m/ D , then we require that the nodes of ! whose immediate predecessor is m are precisely m1; : : : ; mM . The edges rooted at a node correspond in a natural way to the functions in the IFS associated to that node. Let denote the set of all labelled trees which correspond to F . We frequently refer to a labelled tree simply as a tree. A V -variable labelled tree ! 2 is a labelled tree such that, for each level k, there are at most V non-isomorphic subtrees of ! rooted at level k. Let V denote the set of all V -variable labelled trees which correspond to F . 3.3 Approximating fractal sets For E Rn , ! 2 and nodes m D m1 : : : mk 2 !, define E; D E 0 D E; 1/ 1 :::mk Em1 :::mk D fm!.;/ ı fm!.m ı ı fm!.m 1 2 k [ Ek D Em1 :::mk : 1/ .E/; (3.4) jmjDk When we need to make the dependence on ! explicit, we will write Em .!/ or E k .!/, for example. Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 455 V -variable fractals: dimension results For O as in (3.3) we have O Om1 Om1 :::mk ; O O m1 O m1 :::mk ; Om \ On D ; if jmj D jnj; m ¤ n; 1 k O O1 Ok ; O O O ; \ k K ! WD O .!/: (3.5) k0 The set K ! is the fractal set corresponding to !. Even if the open set condition does not apply, we obtain the same set K ! by replacing O by any non-empty compact set E for which M [ fm .E/ E 8 2 ƒ: (3.6) mD1 The set KV of V -variable fractals sets is defined by KV D ¹K ! W ! 2 V º: 4 4.1 The Hausdorff dimension of V -variable fractals The maps ˆ a and F a (See the examples in Figures 3, 4, 5 and 6.) set of all arrays a of the form 2 I a .1/ J a .1; 1/ 6 : :: : aD6 : 4 : a a I .V / J .V; 1/ The index set AV is defined to be the 3 : : : J a .1; M / 7 :: :: 7; : : 5 (4.1) : : : J a .V; M / where I a .v/ 2 ƒ, M is as in (3.1), J a .v; m/ 2 ¹1; : : : ; V º if 1 m MI a .v/ and J a .v; m/ D 0 if MI a .v/ < m M . See also [4, Section 5.3]. The probability distribution PV on AV is obtained as follows. The elements in the first column are chosen independently according to P . The elements J a .v; m/ for 1 m MI a .v/ are chosen independently of one another and of elements in the first column, according to the uniform distribution on ¹1; : : : ; V º. Any remaining elements J a .v; m/ for m > MI a .v/ are set equal to 0. Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 456 M. Barnsley, J. E. Hutchinson and Ö. Stenflo The map ˆaV W V ! V is defined by requiring, for any .!1 ; : : : ; !V / 2 V , that the vth component of ˆa .!1 ; : : : ; !V / 2 V is the labelled tree whose base a node is labelled F I .v/, and whose mth subtree rooted at level one is !J a .v;m/ for 1 m MI a .v/ . Any zeros at the end of each row of a are markers so that all rows have equal length but otherwise play no role. Similarly, for any V -tuple of sets .K1 ; : : : ; KV /, the map F aV W C.Rn / ! Rn / is defined by F a .K1 ; : : : ; KV / 1 0 MI a .V / MI a .1/ [ [ a a fmI .V / .KJ a .V;m/ /A : D@ fmI .1/ .KJ a .1;m/ /; : : : ; mD1 mD1 For a D a0 : : : ak (4.2) 1::: 2 A1 V , assuming (3.6) for some compact E, we define .!1a ; : : : ; !Va / WD lim ˆa0 ı ı ˆak 1 .!10 ; : : : ; !V0 /; k!1 .K1a ; : : : ; KVa / WD lim F a0 ı ı F ak 1 .K10 ; : : : ; KV0 /: (4.3) k!1 The limits exist, and are independent of .!10 ; : : : ; !V0 / and .K10 ; : : : ; KV0 /, respectively. If v 2 ¹1; : : : ; V º, then !va 2 V , and every V -variable labelled tree can be obtained in this manner. However, if V is the set of V -tuples of labelled trees obtained in this manner, then V ¨ .V /V . The probability distribution PV1 on A1 V is defined by selecting the ak in an iid manner according to PV . This induces a probability distribution on V .V /V via (4.3), and thence a probability distribution V on V by projecting in any of the V coordinate directions. Similarly one obtains a probability distribution KV on KV . Both V and KV are independent of choice of projection direction, although the initial distributions are not product distributions. See [4]. 4.2 Approximating Hausdorff measure Assume the open set condition (3.3), and without loss of generality assume that jOj D 1. (See Figure 2 and take O to be the interior of the triangle T .) Suppose ! 2 is a labelled tree with labels from F . Keeping in mind (3.5), we think of the collection of sets ¹O m W m 2 !; jmj D kº as an “efficient” cover of K ! for large k. In order to compute the Hausdorff dimension of K ! , we consider Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM V -variable fractals: dimension results 457 the following quantities for ˛ > 0: r; .!/ WD jOj D 1; rm1 :::mk .!/ WD jO m1 :::mk j !.m1 :::mk !.;/ : : : rm D rm 1 k X S.!; k; ˛/ WD jO m j˛ 1/ for m1 : : : mk 2 !; (4.4) ¹m2!WjmjDkº D X .rm .!//˛ ; noting S.!; 0; ˛/ D 1: ¹m2!WjmjDkº We are interested in the behaviour of S.!; k; ˛/ as k ! 1 since, for large k, it is an approximation to the Hausdorff measure H ˛ .K ! /. 4.3 Flow matrices (See the example in Figure 5.) As remarked in the introduction, there are two ways of representing a V -variable fractal K. One can either use the corresponding labelled tree ! 2 , or one can use a sequence a 2 .AV /1 as in (4.3) which generates a V -tuple containing K as a component. We use the latter in order to study S.!; k; ˛/. For .!1 ; : : : ; !V / 2 V define the following vector of real numbers: S .!1 ; : : : ; !V /; k; ˛ D .S.!1 ; k; ˛/; : : : ; S.!V ; k; ˛// : (4.5) Then S .!1 ; : : : ; !V /; 0; ˛ D 1, the V -vector whose components all equal 1. Proposition 4.1. If .!1 ; : : : ; !V / 2 V , and .!10 ; : : : ; !V0 / D ˆa .!1 ; : : : ; !V /, then ! V a ˛ X X 0 I .v/ S.!v ; k; ˛/ D rm S.!w ; k 1; ˛/: wD1 ¹mWJ a .v;m/Dwº Proof. For any ! 2 it follows from (4.4) that !.;/ rm1 :::mk .!/ D rm rm2 :::mk .! .m1 / /; 1 where ! .m1 / is the labelled subtree of ! rooted at node m1 at level one. From the definition of ˆa .!1 ; : : : ; !V / in Section 4.1, it follows that a I .v/ rm1 :::mk .!v0 / D rm rm2 :::mk .!J a .v;m1 / /: 1 Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 458 M. Barnsley, J. E. Hutchinson and Ö. Stenflo Hence, from (4.4) and the above a ˛ X ˛ I .v/ S.!v0 ; k; ˛/ D rm rm2 :::mk .!J a .v;m1 / / 1 m1 :::mk 2!v0 MI a .v/ D a ˛ I .v/ rm 1 X V X X I a .v/ rm1 ˛ V X X ˛ rm2 :::mk .!w / X ! m2 :::mk 2!w wD1 ¹m1 WJ a .v;m1 /Dwº D ! m2 :::mk 2!J a .v;m1 / m1 D1 D ˛ rm2 :::mk .!J a .v;m1 / / X a .v/ I rm ˛ ! S.!w ; k 1; ˛/: wD1 ¹mWJ a .v;m/Dwº Motivated by this we make the following definition. Definition 4.2. The V V flow matrix M a D M a .˛/ for a 2 AV is defined by a ˛ X I .v/ a ; 1 v; w V: rm Mvw D ¹mWJ a .v;m/Dwº Flow matrices are the main book keeping tool for tracking the size of covers of V -variable fractals. a V Proposition 4.3. Suppose a 2 A1 V , and ! D .!1 ; : : : ; !V / D ! 2 has address a as in (4.3). Then, S .!a ; k; ˛/ D M a0 ı ı M ak i.e. S.!va ; k; ˛/ D V X M a0 ı ı M ak 1 1 vw 1; for 1 v V: wD1 Proof. From Proposition 4.1 and Definition 4.2, for any ! 2 V , S ˆa .!/; k; ˛ D M a S .!; k 1; ˛/: Hence, S ˆa0 ı ı ˆak 1 .!/; k; ˛ D M a0 ı ı M ak DM a0 ı ı M ak 1 S .!; 0; ˛/ 1 1: Since !a is of the form ˆa0 ı ı ˆak 1 .!/ for some !, the result follows. Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM V -variable fractals: dimension results For fixed a D a0 : : : ak 1::: 459 and ! as in Proposition 4.3, we often write Sv .k; ˛/ D S.!v ; k; ˛/: Note that Sv .k; ˛/ depends only on a0 ; : : : ; ak 1, (4.6) and not on aj for j k. 4.4 Computing the Hausdorff dimension P If A is a matrix, we define the norm kAk WD v;w jAvw j. It is easily checked that this norm is submultiplicative, kABk kAk kBk. Main Theorem. Fix F D ¹F W 2 ƒº, a probability distribution P on ƒ, an integer V 1 and a real number ˛ > 0. Under the assumption (3.1), 1 .˛/ WD lim E log kM a0 .˛/ : : : M ak 1 .˛/k k k!1 1 log kM a0 .˛/ : : : M ak 1 .˛/k a.s. D lim k k!1 In particular, the second limit is independent of a D a0 : : : ak 1 : : : for PV1 a.e. a 2 A1 V . The function .˛/ is monotonically decreasing in ˛, and there is a unique d such that .d / D 0. Assuming (3.2), and the open set condition (3.3), dimH .K ! / D d for V a.e. ! 2 V . This theorem is proved in Section 5. For an example see Section 6. Remark 4.4. The limit .˛/ is sometimes called a “Lyapunov exponent”, since kM a0 .˛/ : : : M ak 1 .˛/k grows like e k.˛/ as k ! 1. The fact .˛/ exists and is independent of a, for a.e. a 2 A1 V , is a consequence of the version of the Furstenberg–Kesten theorem [9] in [5, Theorem C, p. 72]. Individual terms k1 log.M a0 .˛/ : : : M ak 1 .˛//vw will not normally converge as k ! 1. In particular, for fixed v and w, .M a0 .˛/ : : : M ak 1 .˛//vw D 0 infinitely often a.s. In fact, if J ak 1 .u; m/ ¤ w for all u and m, which happens with positive probability, then all entries in the w column of M ak 1 .˛/ are zero, and hence all entries in the w column of M a0 .˛/ : : : M ak 1 .˛/ are also zero. However, limk!1 k1 log Sv .k; ˛/ exists a.s. and equals .˛/, for every v 2 ¹1; : : : ; V º, see Lemma 5.5. Remark 4.5. Assuming the open set condition it follows, for random homogeneous fractals K ! (the V D 1 case), that dimH .K ! / is the unique ˛ for which E log M X ˛ .rm / D 0; for a.e. !: (4.7) mD1 Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 460 M. Barnsley, J. E. Hutchinson and Ö. Stenflo See [12] for the direct computation of the dimension in some particular cases. For random recursive fractals K ! , the “V ! 1” case, ! dimH .K / is the unique ˛ for which E M X ˛ .rm / D 1; for a.e. !: (4.8) mD1 See [6, 17, 10, 11]. 5 Proof of the Main Theorem The proof is broken into a number of lemmas. Assumptions We continue the assumptions from the beginning of Section 3, and the notation from Sections 3 and 4. The integer V 1 is fixed and ˛ is non-negative. Define Rmin .˛/ D inf M X ˛ .rm / ; mD1 Rmax .˛/ D sup M X mD1 ˛ .rm / : (5.1) The sequence a D a0 : : : ak 1 : : : 2 AV is chosen according to PV1 as in Section 4.1. The V -tuple of fractal sets corresponding to a is denoted K a , and the corresponding V -tuple of labelled trees is !a . Let K a D .K1a ; : : : ; KVa / D .K1 ; : : : ; KV /; (5.2) !a D .!1a ; : : : ; !Va / D .!1 ; : : : ; !V /: From the next lemma, it follows there is a unique d such that as k ! 1 the product kM a0 .˛/ : : : M ak 1 .˛/k grows exponentially fast to 1 if ˛ < d , and decays exponentially fast to 0 if ˛ > d . This does not imply convergence to a non-zero limit or boundedness of kM a0 .d / : : : M ak 1 .d /k, but it does imply that any infinite growth, or decay to zero, should be slower than exponential. Lemma 5.1. The limit lim E k!1 1 log kM a0 .˛/ : : : M ak 1 .˛/k DW .˛/ k exists. In addition, 1 log kM a0 .˛/ : : : M ak 1 .˛/k D .˛/ k!1 k lim a.s.; Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 461 V -variable fractals: dimension results and in particular is a.s. independent of a. Moreover, log Rmin .˛/ .˛/ log Rmax .˛/: The function is strictly decreasing, Lipschitz, has derivative in the interval Œlog rmin ; log rmax , .0/ > 0 and .˛/ ! 1 as ˛ ! 1. In particular, there is a unique d such that .d / D 0. Proof. The first two claims hold for some .˛/ with 1 .˛/ < 1. This follows from the version of the Furstenberg–Kesten theorem in [5, Theorem C, p. 72] since the ak are chosen in an iid manner. Suppose A and B are square matrices of the same size with non-negative entries. Assume X X Aij ˛2 ; ˇ1 Bij ˇ2 for all i: ˛1 j Let C D AB. Since j P Cij D P ˛1 ˇ1 X j k Aik P j Bkj , Cij ˛2 ˇ2 for all i: j In particular, from Definition 4.2 and (5.1), X k k .˛/ for all v; M a0 .˛/ : : : M ak 1 .˛/ vw Rmax Rmin .˛/ w and so k k .˛/ kM a0 .˛/ : : : M ak 1 .˛/k VRmax .˛/: VRmin Taking logs of both sides, and letting k ! 1, gives the third claim in the lemma. From Definition 4.2, if 0 ˛ < ˇ, k.ˇ ˛/ rmin kM a0 .˛/ : : : M ak 1 .˛/k kM a0 .ˇ/ : : : M ak 1 .ˇ/k k.ˇ rmax ˛/ kM a0 .˛/ : : : M ak 1 .˛/k: Taking logs and letting k ! 1, .ˇ ˛/ log rmin .ˇ/ .˛/ .ˇ ˛/ log rmax : Hence is Lipschitz, differentiable a.e., monotonically decreasing to has derivative in the range log rmin to log rmax . Since 1, and .0/ 2 Œlog Rmin .0/; log Rmax .0/ D Œlog.min M /; log M D log.max M /; it follows that .0/ > 0, and so there exists a unique d such that .d / D 0. Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 462 M. Barnsley, J. E. Hutchinson and Ö. Stenflo Remark 5.2. Bob Scealy [19] has pointed out that one can avoid using the Furstenberg–Kesten theorem and instead use the fact that one has a renewal process, where the renewals are the occurrence of a neck. He has used this idea in his investigation of V -variable fractal graphs. In the next following lemmas, the notion of a “neck” a 2 AV will play an important role, see Figure 6. Figure 6. A neck a with IFSs F , G and H , and with J a .v; m/ D 2 or 0 for all v and m. Definition 5.3. An element a 2 AV is called a neck if all J a .v; m/ are equal for v 2 ¹1; : : : ; V º and if m 2 ¹1; : : : ; MI a .v/º. A neck occurs at level k in a D a0 : : : ak 1 : : : 2 A1 V , or more simply we say that k is a neck, if ak 1 is a neck. Remark 5.4. An element a chosen according to PV is a neck with probability at least V 1 M V . It follows that necks in a sequence a 2 A1 V occur infinitely often a.s. If a neck occurs in a at level k, then all subtrees of .!1a ; : : : ; !Va / rooted at level k are equal. That is, if jmj D jm0 j D k and v; v 0 2 ¹1; : : : ; V º, then mn 2 !v iff m0 n 2 !v0 , and in this case !v .mn/ D !v0 .m0 n/. For the following lemma recall that Sv .k; ˛/ is the sum of the elements in the vth row of M a0 .˛/: : :M ak 1 .˛/. From (4.4) we have that S.!; k; ˛/, for large k, is an approximation to the Hausdorff measure H ˛ .K ! /. See also Remark 4.4. a Lemma 5.5. For a 2 A1 V and ! D ! D .!1 ; : : : ; !V /, lim k!1 1 log Sv .k; ˛/ D .˛/ k a.s. (5.3) Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 463 V -variable fractals: dimension results Proof. Let Sv .k; ˛/ D S.!v ; k; ˛/, as in (4.6). Suppose the address sequence a has a neck at level p, where J ap 1 .v; m/ D u for all v 2 ¹1; : : : ; V º and m 2 ¹1; : : : ; MI ap 1 .v/ º. It follows that all columns of M ap 1 are zero, except for the uth column, and hence the same is true for M a0 ı ı M ap 1 . Suppose A and B are V V matrices such that all columns of A are zero except for the uth column, which we denote by a. Let b be the uth row of B. Then AB D Œ b1 a b2 a : : : bV a : It follows that ! X X X X .AB/vw D av bw D Avw bw : w w w w In particular, the second factor is independent of v. Apply this to X M a0 ı ı M ak 1 vw ; Sv .k; ˛/ D (5.4) w see Proposition 4.3 and (4.6), with A D M a0 ı ı M ap 1 ; B D M ap ı ı M ak 1 : It follows that Sv .k; ˛/ D Sv .p; ˛/ g.k; ˛/; (5.5) where g.k; ˛/ is independent of v, and where Sv .p; ˛/ > 0 for all v. From (5.4), summing over v, X X kM a0 ı ı M ak 1 k D Sv .k; ˛/ D g.k; ˛/ Sv .p; ˛/: v v Hence from Lemma 5.1, .˛/ D lim k!1 1 1 log kM a0 .˛/ : : : M ak 1 .˛/k D lim log g.k; ˛/ k k k!1 a.s. Going back to (5.5), since Sv .p; ˛/ > 0, it follows that lim k!1 1 log Sv .k; ˛/ D .˛/ k a.s. Lemma 5.6. If ˛ > d with d as in Lemma 5.1, then we have H ˛ .Kva / D 0 for v 2 ¹1; : : : ; V º and for a.e. a. In particular, dimH .Kva / d . Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 464 M. Barnsley, J. E. Hutchinson and Ö. Stenflo Proof. We usually drop the reference to a and write Kv for Kva , and !v for !va . Let E be any set such that K1 [ [ KV E, and without loss of generality suppose jEj D 1. For v 2 ¹1; : : : ; V º and m D m1 : : : mk 2 !v , let EvIm1 :::mk D fm!1v .;/ ı fm!2v .m1 / ı ı fm!kv .m1 :::mk 1/ .E/; as in (3.4). Then [ Kv EvIm ; ¹m2!v WjmjDkº and X Sv .k; ˛/ D jEvIm j˛ ; ¹m2!WjmjDkº as in (4.6) and (4.4). Since .˛/ < 0, it follows from (5.3) that 1 lim Sv .k; ˛/ D lim e k. k log Sv .k;˛// D 0 k!1 k!1 a.s. (5.6) Hence H ˛ .Kv / D 0 a.s. Lemma 5.7. Assume F satisfies the open set condition. If ˛ < d , where d is as in Lemma 5.1, then we have H ˛ .Kva / > 0 a.s. for 1 v V . In particular, dimH .Kva / d a.s. Proof. Suppose ˛ < d . As before, Kv D Kva and !v D !va . For a.e. a and each 1 v V , we construct a unit mass measure on Kv such that for some c, .Br .x// cr ˛ if r > 0; x 2 Rn : (5.7) It then follows by the mass distribution principle [7, p. 60] that H ˛ .Kv / > 0, and so dim.Kv / d . a A. Properties of Necks. For a 2 A1 V and k 0, let n .k/ D n.k/ denote the first level k at which a neck occurs. Then we claim 8 > 0 9N > 0 such that 8k n.k/ k N C k a.s.; (5.8) where N will depend on a. To see this, fix > 0 and k > 0, and let Ek D ¹a W na .k/ k > kº: Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 465 V -variable fractals: dimension results It follows from Remark 5.4 that PV1 .Ek / 1 V1 MV k : P Since k1 PV1 .Ek / < 1, it follows from the Borel–Cantelli lemma that, with probability one, Ek occurs for only finitely many k, and so n.k/ k k for all k sufficiently large. Hence for some N depending on a and , n.k/ k N C k for all k. This proves (5.8). B. Construction of and . Suppose a is as in (5.8) and consider the tree !a D .!1 ; : : : ; !V /. For fixed v, a unit measure will first be constructed on the set e !v of infinite paths through !v . For m 2 !v the corresponding cylinder set, a subset of e ! v , is defined by Œm D ¹p 2 e ! v W m pº; where m p means that m is an initial segment of p. The weight function w is defined on cylinder sets by ˛ !v .;/ !v .m1 :::mk w.Œm/ D rvIm WD rm : : : rm 1 k 1/ ˛ : (5.9) We define a unit mass measure on e ! v by setting, if m 2 !v and jmj D k is a neck, w.Œm/ .Œm/ D P : (5.10) 0 ¹w.Œm / W m0 2 !v ; jm0 j D kº The expression for .Œm/ in case jmj is not a neck can be found in (5.15). In order to show this does define a (unit mass) measure on e ! v , first recall that a has necks of arbitrarily large size. We will prove that if k j are both necks, m 2 !v and jmj D k, then satisfies the consistency condition X .Œm/ D ¹.Œn/ W m n 2 !v ; jnj D j º: (5.11) Here, and elsewhere,Sm n means m is an initial segment of the finite sequence n. Note that Œm D ¹Œn W m n 2 !v ; jnj D j º, and this is a union of disjoint sets. It follows from (5.11) that extends to a unit mass measure on the -algebra of subsets of e ! v generated by the cylinder sets Œm for which jmj is a neck. This is just the -algebra generated by all cylinder sets, i.e. the class of Borel sets. In order to prove (5.11), note that if m 2 !v , jmj D k where k is a neck for a, and ms 2 !v , then it follows from (5.9) and Remark 5.4 that ˛ ˛ rvIms D ˛ .s/ rvIm ; (5.12) Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 466 M. Barnsley, J. E. Hutchinson and Ö. Stenflo where .s/ does not depend on either m or v. Suppose now that j k. Then from (5.9) and (5.12), X ¹w.Œn/ W m n 2 !v ; jnj D j º X ˛ (5.13) D rvIm ¹ ˛ .s/ W ms 2 !v ; jsj D j kº ˛ DW .k; j; ˛/ rvIm D .k; j; ˛/ w.Œm/; where does not depend on m or v. Replacing m by m0 and n by n0 , and summing also over m0 , X ¹w.Œn0 / W n0 2 !v ; jn0 j D j º (5.14) X D .k; j; ˛/ ¹w.Œm0 / W m0 2 !v ; jm0 j D kº: Dividing (5.13) by (5.14) and using (5.10) gives (5.11). If m 2 !v with jmj D k, not necessarily a neck, and j n.k/, then X .Œm/ D ¹.Œn/ W m n 2 !v ; jnj D n.k/º P ¹w.Œn/ W m n 2 !v ; jnj D n.k/º D P ¹w.Œn0 / W n0 2 !v ; jn0 j D n.k/º P ˛ ¹rvIn W m n 2 !v ; jnj D n.k/º D P ˛ ¹rvIn0 W n0 2 !v ; jn0 j D n.k/º P ˛ ¹rvIp W m p 2 !v ; jpj D j º D P ˛ ; ¹rvIp0 W p 0 2 !v ; jp 0 j D j º (5.15) using (5.12) in the final equality. Define the map e We ! v ! Kv by e .p1 p2 : : : pk : : :/ D lim fp!1v .;/ ı fp!2v .p1 / ı ı fp!kv .p1 :::pk k!1 1/ .x0 / 2 Kv ; and note that the limit does not depend on x0 . The measure on e ! v projects to the unit mass measure on Kv , defined .A/ D ¹p 2 e !v W e .p/ 2 Aº (5.16) for A a Borel subset of Kv . C. An upper estimate for . Again assume a satisfies (5.8), and consider the corresponding tree !a D .!1 ; : : : ; !V /. Fix v. We show for m D m1 : : : mk 2 !v that ˛ .Œm/ c1 rvIm ; (5.17) for some constant c1 depending on a but not on m. Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM V -variable fractals: dimension results 467 From (5.15), ˛ M n.k/ k ¹r ˛ W m n 2 !v ; jnj D n.k/º rvIm P vIn˛ : Sv .n.k/; ˛/ ¹rvIn0 W n0 2 !v ; jn0 j D n.k/º P .Œm/ D (5.18) To establish this inequality use (5.9) with m there replaced by n, note that each rj˛ 1 and the branching number of !v is bounded by M , and use the expression in (4.4) for Sv .n.k/; ˛/ D S.!v ; n.k/; ˛/. From (5.8), for any > 0, there exists N./ such that M n.k/ k M N./ M k (5.19) for all k. From (5.3), since ˛ < d and so .˛/ > 0, there exists c2 2 R such that for all j , j log Sv .j; ˛/ c2 C .˛/: 2 Hence, 1 1 Sv .n.k/; ˛/ c3 e 2 n.k/.˛/ c3 e 2 k.˛/ (5.20) for some c3 > 0 and all k. 1 Choose 21 .˛/= log M so that e 2 k.˛/ M k , and then choose N D N./. Dividing (5.19) by (5.20), and using (5.18), gives (5.17) with c1 D M N =c3 . D. The estimate for . Fix a 2 A1 V satisfying (5.8), in which case (5.17) holds. Assume the open set condition (3.3) holds with the open set O. Fix x 2 Rn and r > 0. With the measure on Kv as in (5.16), we show by a standard argument, see [15, p. 737] or [7, p. 131], that .Br .x// cr ˛ : (5.21) Here c is independent of x and r, and Br .x/ is the open ball of radius r centred at x. First note that, for each infinite sequence p D m1 m2 : : : mk : : : 2 e !v D e ! av , there is a least k such that rmin r rvIm1 :::mk < r: (5.22) Let Q.r/ D Qa .r/ be the set of all such m D m1 : : : mk . The sets OvIm for m 2 Q.r/ are disjoint from (3.5) and the definition of Q.r/, although the jmj are not necessarily equal. Let Q.x; r/ D Qa .x; r/ be the set of m 2 Q.r/ such that OvIm meets Br .x/. Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 468 M. Barnsley, J. E. Hutchinson and Ö. Stenflo Choose a1 and a2 so that O contains an open ball of radius a1 , and is contained in an open ball of radius a2 . Then the sets OvIm each contain a ball of radius a1 rvIm , and hence of radius a1 rmin r, and they are contained in a ball of radius a2 rvIm , and hence of radius a2 r. It follows by a volume comparison that if q D q.x; r/ is the cardinality of Q.x; r/, since the inner balls are disjoint and are subsets of B.1C2a2 /r .x/, that q.a1 rmin r/n .1 C 2a2 /n r n ; and so q is bounded independently of x and r. Hence, .Br .x// D .Br .x/ \ Kv / D ¹p W .p/ 2 Br .x/ \ Kv º [ ¹Œm W m 2 Q.x; r/º X D ¹.Œm/ W m 2 Q.x; r/º X ˛ c1 ¹rvIm W m 2 Q.x; r/º c1 qr ˛ ; using the definition of Q.x; r/, the disjointedness of the m 2 Q.x; r/ Q.r/, the estimate (5.17) and (5.22). This establishes (5.7) and hence the lemma. 6 Examples For the model problem in the introduction, with F and G each chosen with probability 1=2, it follows from (4.7) that, for random homogeneous Sierpinski triangles, the dimension is d.1/ D 2 log 3=.log 2 C log 3/ 1:226. For the corresponding random recursive case, from (4.8) the dimension is the solution d of d 1 1 d C 21 3 13 D 1, i.e. d.1/ 1:262. For V > 1 we used Maple 10 to 23 2 compute the values of V .˛/ shown in Figure 7. The computed graphs for V > 1 are concave up, although this does not show on the scale used. Subsequent calculation similarly gave .V; d / for 1 V 20 as follows: 1; 6; 11; 16; 1:2262 1:2538 1:2576 1:2590 2; 7; 12; 17; 1:2402 1:2549 1:2580 1:2592 3; 8; 13; 18; 1:2463 1:2557 1:2583 1:2594 4; 9; 14; 19; 1:2500 1:2565 1:2585 1:2596 5; 10; 15; 20; 1:2524 1:2570 1:2588 1:2597 Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 469 V -variable fractals: dimension results 0.03 γ(α) 0.02 0.01 α 1.23 0 d(1) 1.24 d(2) 1.25 d(5) 1.26 1.27 d(infty) –0.01 –0.02 –0.03 –0.04 Figure 7. Graphs of V .˛/ D .˛/ for V D 1; 2; 5 respectively from left to right. Here F D ¹F; Gº and P D ¹1=2; 1=2º as in the introduction. Bibliography [1] M. T. Barlow, Diffusions on fractals, in: Lectures on Probability Theory and Statistics (Saint-Flour, 1995), pp. 1–121, Lecture Notes in Mathematics 1690, SpringerVerlag, Berlin, 1998. [2] M. F. Barnsley, Superfractals, Cambridge University Press, Cambridge, 2006. [3] M. F. Barnsley, J. E. Hutchinson and Ö. Stenflo, A fractal valued random iteration algorithm and fractal hierarchy, Fractals 13, (2005), 111–146. [4] M. F. Barnsley, J. E. Hutchinson and Ö. Stenflo, V -variable fractals: fractals with partial self similarity, Adv. Math. 218 (2008), 2051–2088. [5] J. E. Cohen, Subadditivity, generalized products of random matrices and operations research, SIAM Rev. 30 (1988), no. 1, 69–86. [6] K. Falconer, Random fractals, Proc. Cambridge Philos. Soc. 100 (1986), 559–582. [7] K. Falconer, Fractal Geometry, Mathematical Foundations and Applications, Second Edition, John Wiley & Sons Ltd., Chichester, West Sussex, 2003. [8] K. Falconer, Techniques in Fractal Geometry, John Wiley & Sons Ltd., Chichester, West Sussex, 1997. [9] H. Furstenberg and H. Kesten, Products of random matrices, Ann. Math. Statist. 31 (1960), 457–469. Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM 470 M. Barnsley, J. E. Hutchinson and Ö. Stenflo [10] S. Graf, Statistically self-similar fractals, Probab. Theory Related Fields 74 (1987), 357–392. [11] S. Graf, Random fractals, Rend. Istit. Mat. Univ. Trieste 23 (1991), 81–144. [12] B. M. Hambly, Brownian motion on a homogeneous random fractal, Probab. Theory Related Fields 94 (1992), 1–38. [13] B. M. Hambly, Brownian motion on a random recursive Sierpinski gasket, Ann. Probab. 25 (1997), no. 3, 1059–1102. [14] B. M. Hambly, Fractals and the modelling of self-similarity, in: Stochastic Processes: Modelling and Simulation, pp. 371–406, Handbook of Statistics 21, NorthHolland, Amsterdam, 2003. [15] J. E. Hutchinson, Fractals and self-similarity, Indiana Univ. Math. J. 30 (1981), 713–747. [16] J. Kigami, Analysis on Fractals, Cambridge Tracts in Mathematics 143, Cambridge University Press, Cambridge, 2001. [17] R. D. Mauldin and S. C. Williams, Random recursive constructions: asymptotic geometric and topological properties, Trans. Amer. Math. Soc. 295 (1986), 325–346. [18] P. A. P. Moran, Additive functions of intervals and Hausdorff measure, Proc. Cambridge Philos. Soc. 42 (1946), 15–23. [19] R. Scealy, V -variable fractals and interpolation, Doctoral thesis, Australian National University, Canberra, Australia, 2009. [20] Ö. Stenflo, Markov chains in random environments and random iterated function systems, Trans. Amer. Math. Soc. 353 (2001), 3547–3562. [21] R. S. Strichartz, Differential Equations on Fractals, Princeton University Press, Princeton, NJ, 2006. Received August 20, 2009; revised May 13, 2010. Author information Michael Barnsley, Mathematical Sciences Institute, Australian National University, Canberra, ACT, 0200, Australia. E-mail: [email protected] John E. Hutchinson, Mathematical Sciences Institute, Australian National University, Canberra, ACT, 0200, Australia. E-mail: [email protected] Örjan Stenflo, Department of Mathematics, Uppsala University, 751 05 Uppsala, Sweden. E-mail: [email protected] Brought to you by | Uppsala University Library (Uppsala University Library) Authenticated | 172.16.1.226 Download Date | 5/7/12 7:45 PM